EXAMENSARBETEN I MATEMATIK

EXAMENSARBETEN I MATEMATIK MATEMATISKA INSTITUTIONEN, STOCKHOLMS UNIVERSITET Operator theory in finite-dimensional vector spaces av Kharema Ebshesh 2008 - No 12 MATEMATISKA INSTITUTIONEN, STOCKHOLMS UNIVERSITET, 10691 STOCKHOLM

Operator theory in finite-dimensional vector spaces Kharema Ebshesh Examensarbete i matematik 15 högskolepoäng, påbyggnadskurs Handledare: Andrzej Szulkin och Yishao Zhou 2008

Abstract The theory of linear operators is an extensive area. This thesis is about the linear operators in finite dimensional vector spaces. We study the symmetric, unitary, isometric, and normal operators, and orthogonal projection in the unitary space, the eigenvalue problem and the resolvent. We give a proof of the minimax principle in the end. 1

Acknowledgements I have the pleasure to thank Prof. Andrzej Szulkin and Prof. Yishao Zhou for their help, I enjoyed my work with them, I actually learned a lot during the preparation of this thesis. I would like to thank all the teachers that I had in the Maths Department. I would like to thank Prof. Mikael Passare for inviting me to Sweden. Finally I will not forget my teachers in Libya who supported me during my studies. /Kharema 2

Contents Introduction 4 1 Operators in vector spaces 5 1.1 Vector spaces and adjoint vector spaces............ 5 1.2 Linear operators......................... 12 1.3 Projections............................ 16 1.4 The adjoint operator....................... 17 1.5 The eigenvalue problem..................... 19 1.6 The resolvent........................... 22 2 Operators in unitary spaces 26 2.1 Unitary spaces.......................... 26 2.2 Symmetric operators....................... 29 2.3 Unitary, isometric and normal operators............ 32 2.4 Orthogonal projections...................... 33 2.5 The eigenvalue problem..................... 35 2.6 The minimax principle...................... 36 References 38 3

Introduction This report contains two sections. In Section 1 we study linear operators in finite dimensional vector spaces, where projection, the adjoint operators are introduced. In particular, we study the eigenvalue problem and some properties of the resolvent. In Section 2, different operators in unitary spaces, such as symmetric, unitary, isometric and normal operators are considered. We study the eigenvalue problems for these operators. Finally we prove the minimax principle for eigenvalues. The results in this report are primarily taken from Chapter 1 of [1]. 4

1 Operators in vector spaces 1.1 Vector spaces and adjoint vector spaces Let X be a vector space, and dimx be the dimension of X. In this thesis we shall always assume that dim X <. A subset M of X is a subspace, if M is itself a vector space. We define the codimension of M in X by setting codimm = dim X dim M. Example 1. The set X = C N of all ordered N-tuples u = (ξ i ) = (ξ 1,..., ξ N ) of complex numbers is an N-dimensional vector space. Let dim X = N, if x 1,..., x N are linearly independent, then they are a basis of X, and each u X can be uniquely represented as u = N ξ j x j, (1) j=1 and the scalars ξ j are called the coefficients of u with respect to this basis. Example 2. In C N the N vectors x j = (0,...,0, 1, 0,...,0) with 1 in the j-th place, j = 1,..., N, form a basis (the canonical basis). The coefficients of u = (ξ j ) with respect to the canonical basis are the ξ j themselves. If {x j } is another basis of X, since u = ξ j x j, there is a system of linear relations x k = j γ jk x j, k = 1,..., N. (2) When ξ j,ξ j are coefficients of the same vector u with respect to the bases {x j } and {x j } respectively, they are then related to each other by ξ j = k γ jk ξ k, j = 1,..., N. (3) The inverse transformations to (2) and (3) are x j = k ˆγ kj x k, ξ k = j ˆγ kj ξ j, (4) where (ˆγ jk ) is the inverse of the matrix (γ jk ) : 5

ˆγ ji γ ik = i i γ jiˆγ ik = δ jk = { 1, j = k 0, j k. Let M 1, M 2 be subspaces of X, we define the subspace M 1 + M 2 by (5) M 1 + M 2 = {x 1 + x 2 : x 1 M 1, x 2 M 2 }. (6) Theorem 1. If M 1, M 2 are subspaces of X, then dim(m 1 + M 2 ) = dim M 1 + dimm 2 dim(m 1 M 2 ). (7) Proof. We can see that M 1 M 2 is subspace of both M 1 and M 2. Suppose that dimm 1 = m 1, dimm 2 = m 2 and dim(m 1 M 2 ) = m. We have to prove that dim(m 1 + M 2 ) = m 1 + m 2 m. Suppose that {x 1,..., x m } is a basis of M 1 M 2. We can extend this basis to a basis of M 1 and to a basis of M 2. Let for example { x 1,..., x m, x (1) 1 },..., x(1) m 1 m, { x 1,..., x m, x (2) 1 },..., x(2) m 2 m be two bases { of M 1 and M 2 respectively. } Let B = x 1,..., x m, x (1) 1,..., x(1) m 1 m, x(2) 1,..., x(2) m 2 m. Then B generates M 1 + M 2. Now we will show that the vectors in B are linearly independent. Suppose that α 1 x 1 +...+α m x m +β 1 x (1) 1 +...+β m 1 mx (1) m 1 m +γ 1x (2) 1 +...+γ m 2 mx (2) m 2 m = 0. (8) Let we have also u = α 1 x 1 +... + α m x m + β 1 x (1) 1 +... + β m1 mx (1) m 1 m, (9) u = γ 1 x (2) 1... γ m2 mx (2) m 2 m, (10) 6

thus u M 1 and u M 2 by (9) and (10). Hence u M 1 M 2. Therefore there are ζ i, i = 1,..., m such that u = ζ i x i, and { Since ζ 1 x 1 +... + ζ m x m + γ 1 x (2) 1 +... + γ m2 mx (2) m 2 m = 0. x 1,..., x m, x (2) 1 },..., x(2) m 2 m is a basis of M 2, then γ 1 = 0,..., γ m2 m = 0. Substituting this in (8), we obtain α 1 x 1 +... + α m x m + β 1 x (1) 1 +... + β m1 mx (1) m 1 m = 0, thus α 1 = 0,..., α m = 0, β 1 = 0,..., β m1 m = 0. Hence B is a basis of M 1 +M 2. Since B has m 1 +m 2 m vectors, we get the required result. Definition 1. (Direct sum) Let M 1,..., M s be subspaces of X. X is a direct sum of them if each u X has a unique expression of the form u = j u j, u j M j, j = 1,..., s. (11) Then we write X = M 1... M s. Proposition 1. If X = M 1... M s, then 1. X = M 1 +... + M s. 2. M i Mj = {0}, where i j. 3. dimx = j dimm j. Proof. If X = M 1... M s, then each u X can be uniquely expressed as u = u 1 +... + u s, where u j M j. Hence X = M 1 +... + M s. Now let u M i M j, where i j, then u = 0 1 +... + 0 i 1 + u i + 0 i+1 +... + 0 j +... + 0 s, and u = 0 1 +... + 0 i +... + 0 j 1 + u j + 0 j+1 +... + 0 s. Since the expression (11) for u is unique, u i = u j = 0 and hence u = 0. To show the last equality in the proposition we assume that dim X = N, dimm j = m j, and we have to show that N = m j. 7

Let x 1 i,...,xs i, i = 1,..., m j be bases of M j, j = 1,..., s respectively. Suppose that mj m 1 i=1 m 2 m s α 1i x 1 i + α 2i x 2 i +... + α si x s i = 0. (12) i=1 i=1 α jix j i M j, and since X is the direct sum of M j, the representation (12) is unique, and so for j = 1,..., s we have m j i=1 α jix j i = 0. Since {xj i } are linearly independent, α ji = 0 for i = 1,..., m j, j = 1,..., s. Hence x j i, i = 1,..., m j, j = 1,..., s are linearly independent. Now suppose u X, u = u 1 +u 2 +...+u s, u j M j. Since u j is expressed in a unique way as a linear combination of x j i, i = 1,..., m j, then u is expressed in a unique way as a linear combination of x j i, j = 1,..., s. Hence xj i, j = 1,..., s is a basis of X. Thus i=1 dimx = j dimm j. Definition 2. We call u a norm of u X, if (i) u 0 for all u X and u = 0 iff u = 0. (ii) αu = α u for all α C, u X. (iii) u + v u + v for all u, v X. Example 3. u = max ξ j, j u = j u = j ξ j 1 2 ξ j 2, where ξ j are coefficients of u with respect to the basis {x j } in X, are three different norms. 8

We show only that the last expression is a norm. Let ξ j,η j be the coefficients of u, v respectively, then - u = ( ξ j 2)1 2 0 and u = 0 iff ξ j = 0, j. - αu = ( αξ j 2)1 2 = ( α 2 ξ j 2)1 2 = α ( ξ j 2)1 2. -It follows from the Schwarz inequality (see e.g. [2]) that ( ξ j η j ) 2 ξ j 2 η j 2, therefore 2 ( ( ξ j η j 2 ξj 2)1 2 ηj 2)1 2. Adding ξ j 2 + η j 2 to both sides gives ( ξj 2 + 2 ξ j η j + η j 2 ) = ξ j 2 + 2 ξ j η j + η j 2 ξ j 2 + 2( ξ j 2 ) 1 2 ( ηj 2 ) 1 2 + ηj 2 = [( ξ j 2 ) 1 2 + ( ηj 2 ) 1 2 ] 2. Hence that is ( ξ j + η j 2 ) 1 2 ( ξj 2 ) 1 2 + ( ηj 2 ) 1 2, u + v u + v. -The adjoint space Definition 3. (Linear forms and semilinear forms): A complex-valued function f[u] defined for u X is called a linear form if f[αu + βv] = αf[u] + βf[v] (13) for all u, v X, and all scalars α, β, and a semilinear form if f[αu + βv] = ᾱf[u] + βf[v]. (14) Example 4. Let x 1,..., x N be a fixed basis in X. It follows from (13) that a linear form on X can be expressed in the form f[u] = N α j ξ j, where u = (ξ j ) and f[x j ] = α j, j=1 9

and similarly by (14) a semilinear form on X can be expressed in the form f[u] = j α j ξj. f[u] is a semilinear form if and only if f[u] is a linear form. Definition 4. The set of all semilinear forms on X is a vector space, called the adjoint (or conjugate) space of X and is denoted by X. Let us denote f[u] by (f, u) where f is a semilinear form. It follows from the definition that (f, u) is linear in f and semilinear in u: (αf + βg, u) = α(f, u) + β(g, u), (15) (f, αu + βv) = ᾱ(f, u) + β(f, v). (16) Example 5. For X = C N, X may be regarded as the set of all row vectors f = (α j ) whereas X is the set of all column vectors u = (ξ j ), and (f, u) = α j ξj. The adjoint basis The principal content in this part is the following theorem: Theorem 2. Suppose {x j } is a basis of X, and let e 1,..., e N be vectors in X defined by { 1, j = k (e j, x k ) = δ jk = 0, j k, (17) then {e j } is a basis of X. Proof. First we show that e j satisfying (17) exist. Define e j, j = 1,..., N by (e j, u) = ξ j. Then this corresponds to α k = δ jk, k = 1,..., N, in the formula above. Next we shall show that {e j } generate X. Let f X, and suppose (f, x 1 ) = α 1, (f, x 2 ) = α 2,...,(f, x N ) = α N. Put g = α j e j, then (g, x 1 ) = ( α j e j, x 1 ) = α 1 (e 1, x 1 ) + α 2 (e 2, x 1 ) +... + α N (e N, x 1 ) = α 1, and similarly for j = 2,..., N,(g, x j ) = α j so that f(x j ) = g(x j ), j = 1,..., N. Since f, g are equal on vectors of the basis {x j }, then f = g = α 1 e 1 +... + α N e N, 10

i.e., f is a linear combination of e 1,..., e N. It remains to show e 1,..., e N are linearly independent. Suppose that α 1 e 1 +... + α N e N = 0. Then 0 = (0, x 1 ) = (α 1 e 1 +... + α N e N, x 1 ) = α 1 (e 1, x 1 ) +... + α N (e N, x 1 ) = α 1, and similarly for k = 2,..., N, so that we have α 1 =... = α N = 0. Thus e 1,..., e N are linearly independent. Hence {e j } is a basis of X. Let {x j } be a basis of X, and let {e 1,..., e N }, {e 1,..., e N } be vectors in X satisfying (17). Then by Theorem 2 {e j } and {e j } are bases of X and so that e i = N j=1 α (i) j e j, i = 1,..., N, (e 1, x 1 ) = α (1) 1 (e 1, x 1 ) + α (1) 2 (e 2, x 1 ) +... + α (1) N (e N, x 1 ) = α (1) 1 1 + α (1) 2 0 +... + α (1) 2 0 = α (1) 1, and (e 1, x 2) = α (1) 2,...,(e 1, x N) = α (1) N. Hence α(1) j = δ j1 and e 1 = e 1. Similarly for j = 2,..., N we obtain e j = e j. Hence the basis {e j } of X that satisfies (17) is unique. It is called the basis adjoint to the basis {x j } of X. Theorem 2 shows that dimx = dimx. (18) Let {x j } and {x j } be two bases of X related to each other by (2). Then the corresponding adjoint bases {e j } and {e j } of X are related to each other by the formulas e j = k γ jk e k, e k = j ˆγ kj e j. (19) Furthermore we have γ jk = (e j, x k ), ˆγkj = (e k, x j). (20) Definition 5. Let f X. The norm f is defined by (f, u) f = sup = sup (f, u). (21) 0 u X u u =1 11

1.2 Linear operators Definition 6. Let X, Y be two vector spaces. A function T that sends every vector u of X into a vector v = Tu of Y is called a linear transformation or a linear operator on X to Y if for all u 1, u 2 X and all scalars α 1, α 2. T(α 1 u 1 + α 2 u 2 ) = α 1 Tu 1 + α 2 Tu 2 (22) If Y = X we say T is a linear operator in X. If M is a subspace of X, then T(M) is a subspace of Y, the subspace T(X) of Y is called the range of T and is denoted by R(T), dim(r(t)) is called the rank of T. The codimension of R(T) with respect to Y is called the deficiency of T and is denoted by def T, hence rank T + def T = dimy. (23) The set of all u X such that Tu = 0 is a subspace of X and is called the kernel or null space of T and is denoted by N(T). dim(n(t)) is denoted by nul T, and we have rank T + nul T = dimx. (24) If both nul T and def T are zero, then T is one-to-one. In this case the inverse operator T 1 is defined. Let {x k } be a basis of X. Each u X has the expansion (1) so that Tu = N ξ k Tx k, N = dimx. (25) k=1 Thus an operator T on X to Y is determined by giving the values of Tx k, k = 1,..., N. If {y j } is a basis of Y, each Tx k has the expansion Tx k = M τ jk y j, M = dimy. (26) j=1 Substituting (26) into (25), the coefficients η j of v = Tu are given by η j = N τ jk ξ k, j = 1,..., M. (27) k=1 12

In this way an operator T on X to Y is represented by an M N matrix (τ jk ) with respect to the bases {x k },{y j } of X, Y respectively. When (τ jk ) is the matrix representing the same operator T with respect to a new pair of bases {x k }, {y j }, we can find the relationship between the matrices (τ jk ) and (τ jk) by combining (26) and a similar expression for Tx k in terms of {y j } with the formulas (2),(4): Tx k = T( h ˆγ hk x h ) = h = h = i = i = j = j = j ˆγ hk Tx h M ˆγ hk τ ih y i i=1 τ ihˆγ hk y i h τ ihˆγ hk h j γ jiy j γ jiτ ihˆγ hk y j i h γ jiτ ihˆγ hk y j i,h τ jk y j, where x h = k γ kh x k and x k = h ˆγ hk x h, y i = j γ jiy j and y j = i γ ijy i. Thus the matrix τ jk is the product of three matrices (γ jk ), (τ jk) and (ˆγ jk ), (τ jk ) = (γ jk )(τ jk)(ˆγ jk ). (28) Thus when T is an operator on X to itself det (τ jk ) and the trace of (τ jk ), i.e., τ jj are determined by the operator T itself. More precisely, we shall show det (τ jk ) and trace (τ jk ) are the same for each choice of the basis for X. (28) becomes (τ jk ) = (γ jk)(τ jk )(ˆγ jk ). (29) We show tr (γτ) = tr (τγ). Let γτ = (a jk ) and τγ = b jk, then tr (γτ) = a jj = γ jk τ kj = τ kj γ jk = b kk = tr (τγ). Since (ˆγ jk ) is j j k k j k 13

the inverse of the matrix (γ jk ), and we know that det(γτˆγ) = det(γ)det(τ)det(ˆγ), we have det(τ ) = det(τ), and tr (τ ) = tr (γτˆγ) = tr (τγˆγ) = tr (τ). (30) Example 6. If {f j } is the basis of Y adjoint to {y j } then τ jk = (f j, Tx k ). (31) Proof. Since Tx k = i τ ik y i, then (f j, Tx k ) = (f j, i τ ik y i ) = i τ ik (f j, y i ) = i τ ik δ ij = τ jk. Example 7. Let {x j } and {e j } be the bases of X and X, respectively, which are adjoint to each other. If T is an operator on X to itself, we have tr T = j (e j, Tx j ). (32) Proof. Similarly as in the last example, Tx j = i τ ij x i, therefore (e j, Tx j ) = j j = j = j = j (e j, τ ij x i ) i ( τ ij (e j, x i )) i τ ij δ ij i τ jj. 14

If T and S are two linear operators on X to Y, then we define: (αs + βt)u = α(su) + β(tu). If S maps X to Y and T maps Y to Z, then we set (TS)u = T(Su). Example 8. 1. rank (S + T) rank S + rank T. 2. rank (T S) max(rank T, rank S). Proof. (1) Let R(S) = M 1, R(T) = M 2, R(S + T) = M. Since each v M can be expressed in the form v = v 1 + v 2, v 1 M 1, v 2 M 2, M = M 1 + M 2, thus dimm = dimm 1 +dimm 2 dim(m 1 M 2 ) dim M 1 +dimm 2. Hence rank (S + T) rank (S) + rank (T). (2) Let S : X Y, T : R(S) Z; TS : X Z. Then we have rank S + nul S = dim X and rank (TS) + nul (TS) = dimx. Since nul S nul TS, rank TS rank S. Let T : Y Z, TS : X Z. Since T : Y Z, rank T +def T = dimz, and rank (TS)+def (TS) = dimz. Since def T def (TS), rank (TS) rank T. Thus rank (TS) max(rank T, rank S). Let us denote by L(X, Y ) the set of all operators on X to Y. It is a vector space. Let L(X) = L(X, X), then we have: T0 = 0T = 0. T1 = 1T = T; (1 is the identity operator). T m T n = T m+n, (T m ) n = T mn, m, n = 0, 1,...,. If S, T L(X) are nonsingular, then T 1 T = 1, T n = (T 1 ) n, and (TS) 1 = S 1 T 1. For any polynomial P(z) = α 0 + α 1 z +... + α n z n in the indeterminate z, we define the operator P(T) = α 0 + α 1 T +... + α n T n. Example 9. If S L(X, Y ) and T L(Y, X), then ST L(Y ) and TS L(X). 15

1.3 Projections Let M, N be two complementary subspaces of X, X = M N. Thus each u X can be uniquely expressed in the form u = u + u, u M, u N. u is called the projecion of u on M along N. If v = v + v, then αu + βv has the projection αu + βv on M along N. If we set u = Pu, it follows that P is a linear operator in X called the projection operator or projection on M along N. 1 P is the projection on N along M, and we have Pu = u if and only if u M, Pu = 0 if and only if u N, that is R(P) = N(1 P) = M, N(P) = R(1 P) = N. Furthermore, PPu = Pu, that is P is idempotent, i.e., P 2 = P. Remark 1. Any idempotent operator P is a projection. To show, let M = R(P) and N = R(1 P). If u M, there is u such that Pu = u and therefore Pu = P 2 u = Pu = u. Similarly if u N. Now let u M N. Then u = Pu = 0. So M N = {0}. Thus each u X has the expression u = u + u with u = Pu M and u = (1 P)u N, proving that P is the projection on M along N. Example 10. If P is a projection, then we have tr P = dim R(P). Proof. Since P is an operator in X, it can be represented by an n n (n = dimx) matrix (τ jk ) with respect to the basis {x j } of X, and Pu = u when u M. This basis can be chosen so that x 1,..., x m M and x m+1,..., x n N, where N = (1 P)X. Then P(α 1 x 1 +... + α n x n ) = α 1 x 1 +... + α m x m. Hence (τ jk ) is diagonal with τ 11 =... = τ mm = 1, τ m+1,m+1 =... = τ nn = 0. So tr P = m = dim R(P). In general, if X = M 1... M s, then each u X can be uniquely expressed in the form u = u 1 +... + u s, u j M j, j = 1,..., s. Then the operator P j defined by P j u = u j M j, is a projection on M j along N j = M 1... M j 1 M j+1... M s. And we have Pj = 1, (33) 16

for ( j P j )( i u i ) = j P j ( i u i ) = P j u j = u j, and P k P j = δ jk P j, (34) because P k P j ( i u i ) = P k u j = δ kj u j = δ kj P j u. Note that, if we have (33) and (34), then X is the direct sum of subspaces R(P j ). To show that, let M j = R(P j ). For u X, by (33) we have u = Pj u = u j M 1 +... + M s. Moreover M i M j = {0} for i j, because by (34) if u M i M j, then u = P i u 1 = P j u 2 and u = P i u 1 = Pi 2u 1 = P i P j u 2 = 0. Hence X = M 1... M s. Since P j is idempotent, it follows from Remark 1 that P j is the projection on M j along N j. 1.4 The adjoint operator Definition 7. Let T L(X, Y ), a function T on Y to X is called the adjoint operator of T if : (T g, u) = (g, Tu), g Y, u X. (35) Then (T (α 1 g 1 +α 2 g 2 ), u) = (α 1 g 1 +α 2 g 2, Tu) = α 1 (g 1, Tu)+α 2 (g 2, Tu) = α 1 (T g 1, u)+α 2 (T g 2, u) so that T (α 1 g 1 +α 2 g 2 ) = α 1 T g 1 +α 2 T g 2. Therefore T is a linear operator on Y to X, that is, T L(Y, X ). The operation * has the following properties: 1. (αs + βt) = ᾱs + βt, for S, T L(X, Y ), and α, β C. 2. (TS) = S T, for T L(Y, Z) and S L(X, Y ). Note that S L(Y, X ) and T L(Z, Y ) so that S T L(Z, X ). Then ((TS) h, u) = (h, TSu) = (T h, Su) = (S T h, u), h Z, u X. Hence 2. holds. Example 11. If T L(X), we have 0 = 0, 1 = 1. If {x k }, {y j } are bases in X, Y respectively, and T L(X, Y ) is represented by a matrix (τ jk ) in these bases, and {e k }, {f j } are the adjoint bases of X, Y respectively, the operator T L(Y, X ) can be represented by a matrix (τ kj ). These matrices are given by τ jk = (f j, Tx k ) according to (31) and τ kj = (T f j, x k ) = (f j, Tx k ) (see the argument below), thus τ kj = τ jk, k = 1,..., N = dim X, j = 1,..., M = dimy. (36) 17

To show that (T f j, x k ) = τ kj, we first write T f j = i τ ijf i and then compute: (T f j, x k ) = ( i τ ijf i, x k ) = i τ ij(f i, x k ) = i τ ijδ ik = τ kj. Example 12. If T L(X), we have det T = det T, tr T = tr T (37) and (T ) 1 = (T 1 ). (38) Since det (τ jk ) and tr (τ jk ) are the same for each choice of the basis for X and similarly with (τkj ), (37) is satisfied according to (36). To prove (38) note that T (T 1 ) = (T 1 T) = 1 = 1. Definition 8. (Norm of T) The norm of T is defined by Example 13. Proof. Tu T = sup 0 u X u = sup Tu, T L(X, Y ). (39) u =1 T = (f, Tu) sup 0 u X f u = sup 0 f Y u =1 f =1 (f, Tu). (40) We first prove that the expression for T given by (40) is a norm. - T = sup - αt = sup u =1 f =1 - T+S = sup u =1 f =1 (f, Su) ) sup u =1 f =1 in (40) is a norm. u =1 f =1 (f, Tu) 0, and = 0 iff T = 0. (f, αtu) = sup u =1 f =1 (f, Tu+Su) = sup (f, Tu) + sup u =1 f =1 α(f, Tu) = α sup u =1 f =1 u =1 f =1 (f, Tu) = α T. (f, Tu)+(f, Su) sup ( (f, Tu) + u =1 f =1 (f, Su) = T + S. Hence T defined 18

We have to show that (39) and (40) are equivalent. To see this we recall (f, u) (21). Note that (f, u) f u. This implies that u = sup = 0 f X f sup (f, u) (see Section I.2.5 in [1]). It follows that Tu = sup (f, Tu) f =1 f =1 Tu = sup (f, Tu). and T = sup u =1 u =1 f =1 Tu Tu Since T = sup, then u =1 u u TSu T Su T S u, thus for T L(Y, Z) and S L(X, Y ). If T L(X, Y ), then T (Y, X ) and T, so Tu T u. Hence TS T S (41) T = T. (42) This follows from (40) according to which T = sup (T f, u) = sup (f, Tu) = T where u X, u = 1 and f X, f = 1. 1.5 The eigenvalue problem Definition 9. Let T L(X). A complex number λ is called an eigenvalue of T if there is a non-zero vector u X such that Tu = λu. (43) u is called an eigenvector of T belonging to the eigenvalue λ. The set N λ of all u X such that Tu = λu is a subspace of X called the eigenspace of T for the eigenvalue λ and dimn λ is called the multiplicity of λ. Example 14. λ is an eigenvalue of T if and only if λ ξ is an eigenvalue of T ξ. Proposition 2. The eigenvectors of T belonging to different eigenvalues are linearly independent. Proof. To prove that we will use induction. We shall first show that any two eigenvectors of T belonging to different 19

eigenvalues are linearly independent. Assume next this is true for k eigenvectors, and we shall prove it for k +1 eigenvectors. Let Tu 1 = λ 1 u 1, Tu 2 = λ 2 u 2, λ 1 λ 2, λ 1 0 and α 1 u 1 + α 2 u 2 = 0, we have α 1 λ 1 u 1 + α 2 λ 2 u 2 = 0. By multiplying the first equation by λ 1 and subtracting it from the second we obtain (λ 1 λ 2 )α 2 u 2 = 0. Hence α 2 u 2 = 0, but u 2 0, so α 2 = 0. Now α 1 u 1 = 0 which implies α 1 = 0 since u 1 0. Hence α 1 = α 2 = 0, that is u 1, u 2 are linearly independent. Now assume Tu 1 = λ 1 u 1,..., Tu k = λ k u k, λ i λ j for i j and u 1,..., u k are linearly independent. We shall show that u 1,..., u k, u k+1 are linearly independent where Tu k+1 = λ k+1 u k+1. We have two cases: λ k+1 = 0 and λ k+1 0. If λ k+1 = 0, then we have λ i 0, i = 1,..., k. If u 1,..., u k, u k+1 are linearly dependent then u 1 = α 2 u 2 +... + α k u k + α k+1 u k+1, and λ 1 u 1 = Tu 1 = α 2 λ 2 u 2 +... + α k λ k u k, thus u 1,..., u k are linearly dependent, and this contradicts our assumption. If λ k+1 0 suppose u k+1 = α 1 u 1 +... + α k u k, where α 1 0, then we have λ k+1 u k+1 = Tu k+1 = α 1 λ 1 u 1 +... + α k λ k u k = λ k+1 ( α 1λ 1 λ k+1 u 1 +... + α kλ k λ k+1 u k ). Since u 1,..., u k are linearly independent, we obtain λ 1 = λ k+1,..., λ k = λ k+1, and this is also a contradiction. It follows from this proposition that there are at most N eigenvalues of T, where N is the dimension of X. 20

Proposition 3. lim T n 1 n exists and is equal to inf T n 1 n. n n=1,2,... Proof. It follows from (41) that T m T n T m T n, T n T n, m, n = 0, 1,... (44) Set a n = log T n, what is to be proved is that The inequality (44) gives a n lim n n = inf a n n=1,2, n. (45) a m+n a m + a n. Let n = mq + r, where q, r are nonnegative integers with 0 r < m, then the last inequality gives a n a mq + a r. Let a mq = log T mq. By (44) a mq log T m q = q log T m = qa m, hence and Therefore Since lim n sup q n = 1 m a n qa m + a r a n n q n a m + 1 n a r. lim sup a n n n lim sup q n n a m + lim sup a r n n. Since this holds for all fixed m, and lim n sup a r n = 0, lim sup a n n n a m m. Obviously hence lim sup a n n n inf a m m. lim sup a n n n inf a m m, lim sup a n n n = inf a n n=1,2,... n. 21

Now we define: Definition 10. (Spectral radius of T) sprt =lim T n 1 n =inf T n 1 n. 1.6 The resolvent Let T L(X) and consider the equation (T ξ)u = v, where ξ is a given complex number, v X is given and u X is to be found. This equation has a solution u for every v if and only if T ξ is nonsingular, that is ξ is different from any eigenvalue λ k of T. Then the inverse (T ξ) 1 exists and the solution u is given by The operator u = (T ξ) 1 v. R(ξ) = R(ξ, T) = (T ξ) 1 (46) is called the resolvent of T. The complementary set of the spectrum Σ(T) is called the resolvent set of T and will be denoted by P(T). The resolvent R(ξ) is thus defined for ξ P(T). Example 15. R(ξ) commutes with T. And R(ξ) has exactly the eigenvalues (λ h ξ) 1 where λ h are eigenvalues of T. We show that R(ξ) commutes with T: Hence T = T 1 = T(T ξ)(t ξ) 1 = (T ξ)t(t ξ) 1. (T ξ) 1 T = (T ξ) 1 (T ξ)t(t ξ) 1 = T(T ξ) 1. Now we show that if λ is an eigenvalue of T, that is Tu = λu, then (λ ξ) 1 is an eigenvalue of R(ξ). Clearly we have (T ξ)u = (λ ξ)u, 22

or equivalently (λ ξ) 1 (T ξ)u = u. Then i.e. Hence (T ξ) 1 ((λ ξ) 1 (T ξ)u) = (T ξ) 1 u, (λ ξ) 1 ((T ξ) 1 (T ξ)u) = (T ξ) 1 u. (λ ξ) 1 u = (T ξ) 1 u. since Note that the resolvent satisfies the (first) resolvent equation R(ξ 1 ) R(ξ 2 ) = (ξ 1 ξ 2 )R(ξ 1 )R(ξ 2 ) (47) (T ξ 1 ) 1 (T ξ 2 ) 1 = (T ξ 1 ) 1 (T ξ 2 ) 1 (T ξ 2 )(T ξ 1 ) [(T ξ 1 ) 1 (T ξ 2 ) 1 ] = (T ξ 1 ) 1 (T ξ 2 ) 1 [(T ξ 2 ) (T ξ 1 )] = (ξ 1 ξ 2 )(T ξ 1 ) 1 (T ξ 2 ) 1 = (ξ 1 ξ 2 )R(ξ 1 )R(ξ 2 ). Here we have used the identity (T ξ 2 )(T ξ 1 ) = (T ξ 1 )(T ξ 2 ). We shall show that for each ξ 0 P(T) R(ξ) is holomorphic in some disk around ξ 0. Proposition 4. R(ξ) = (ξ ξ 0 ) n R(ξ 0 ) n+1 is absolutely convergent for ξ ξ 0 < (spr R(ξ 0 )) 1 where ξ 0 is a given complex number. To prove this we have to study the following Lemmas. Lemma 1. (Neumann series) The series T < 1. Moreover, (1 T) 1 = T n is absolutely convergent if n=0 T n, and (1 T) 1 (1 T ) 1, where T L(X). (48) n=0 23

Proof. This series is absolutely convergent because lim T n 1 n T < 1. n Set T n = S, then n=0 TS = = = n=0 n=1 T n+1 T n T n 1 n=0 = S 1, so that TS = ST = S 1. Hence (1 T)S = S(1 T) = 1 and S = (1 T) 1. Now we have (1 T) 1 = T n T n T n = (1 T ) 1. Lemma 2. The series (48) is absolutely convergent if T m < 1 for some positive integer m, or equivalently, if spr T < 1, and the sum is again equal to (1 T) 1 (see proof of Proposition 3). Proof. Since spr T = inf n=1,2,... T n 1 n n=0 n=0 n=0 = lim n T n 1 n and T m 1 m < 1, it follows that lim n T n 1 n < 1 and the series is absolutely convergent. The proof that the sum is equal to (1 T) 1 is the same as above (see the proof of Proposition 3). Lemma 3. S(t) = (1 tt) 1 = t n T n, (49) where t is a complex number. The convergence radius r of (49) is exactly equal to 1/spr T. Proof. By Lemma 2, (49) holds if spr(tt) < 1, i.e., t < 1/sprT, so the convergence radius r 1/sprT. If t > 1/sprT, then spr(tt) > 1, so lim n tn T n 1 n > 1 and the series diverges. Hence r = 1/sprT. n=0 24

Now we can complete the proof of Proposition 4. By (47), with ξ 1 = ξ and ξ 2 = ξ 0, we have R(ξ) = R(ξ 0 )(1 (ξ ξ 0 )R(ξ 0 )) 1. Let tt = (ξ ξ 0 )R(ξ 0 ) in Lemma 3. Then R(ξ) = By Proposition 4 we obtain: (ξ ξ 0 ) n R(ξ 0 ) n+1. n=0 Proposition 5. R(ξ) is holomorphic at ξ 0 in the disk ξ ξ 0 < (spr T) 1. Proposition 6. R(ξ) is holomorphic at. Proof. R(ξ) has the expansion R(ξ) = ξ 1 (1 ξ 1 T) 1 = ξ n 1 T n, (50) which is convergent if and only if ξ > spr T, thus R(ξ) is holomorphic at infinity. Example 16. R(ξ) ( ξ T ) 1 and R(ξ) + ξ 1 ξ 1 ( ξ T ) 1 T, for ξ > T. n=0 Proof. It follows from (50) that R(ξ) = ξ n 1 T n = ξ 1 (T/ξ) n ξ 1 ξ n T n ξ 1 ξ n T n = ξ 1 (1 ξ 1 T ) 1 = ( ξ T ) 1, 25

and R(ξ) + ξ 1 = = ξ n 1 T n + ξ 1 n=0 ξ n 1 T n n=1 ξ 1 ξ n T n n=1 = ξ 1 [(1 ξ 1 T ) 1 1] 1 = ξ T 1 ξ ξ ξ + T = ξ ( ξ T ) = ξ 1 ( ξ T ) 1 T. The spectrum Σ(T) is never empty; T has at least one eigenvalue. Otherwise R(ξ) would be an entire function such that R(ξ) 0 for ξ, then we must have R(ξ) = 0 by Liouville s theorem (see [3]). But this results in the contradiction that 1 = (T ξ)r(ξ) = 0. We can see that each eigenvalue of T is a singularity of the analytic function R(ξ). Since there is at least one singularity of R(ξ) on the convergence circle ξ = spr T according to (50), spr T coincides with the largest (in absolute value) eigenvalue of T: spr T = max λ h. (51) This shows that spr T is independent of the norm used in its definition. 2 Operators in unitary spaces 2.1 Unitary spaces A normed space X is a special case of a linear metric space in which the distance between any two points is defined by u v, where u and v belong to X. 26

Definition 11. (The complex inner product) Let u, v X, and let (u, v) be a complex number, then we say that the function (, ) is a complex inner product if it satisfies: -(αu 1 + βu 2, v) = α(u 1, v) + β(u 2, v). -(u, v) = (v, u). -(u, u) > 0, if u 0. From the second condition in the last definition we can obtain (u, kv) = k(u, v), since (u, kv) = (kv, u) = k(v, u) = k(u, v). Definition 12. A normed space H is called a unitary space if an inner product (u, v) is defined for all vectors u, v H. Definition 13. In a unitary space the function is a norm which is called the unitary norm. u = (u, u) 1 2 (52) We shall show the conditions in the Definition 2: -the first condition follows directly from the definition of the inner product. - αu = (αu, αu) 1 2 = [α(u, αu)] 1 2 = [αᾱ(u, u)] 1 2 = α u. - u + v = (u + v, u + v) 1 2 = [(u, u + v) + (v, u + v)] 1 2 = [(u, u) + (u, v) + (v, u) + (v, v)] 1 2 [(u, u) + (u, v) + (v, u) + (v, v)] 1 2. By the Schwarz inequality (u, v) u v, (53) hence u+v [ u 2 + (u, v) + (v, u) + v 2 ] 1 2 [ u 2 +2 u v + v 2 ] 1 2 = u + v. Example 17. For numerical vectors u = (ξ 1,..., ξ N ) and v = (η 1,..., η N ) set (u, v) = ξ j η j, u 2 = ξ j 2, with this inner product the space C N becomes a unitary space. 27

Remark 2. A characteristic property of a unitary space H is that the adjoint space H can be identified with H itself. To show that, let f, u H, we have the form (f, u) is linear in f and semilinear in u by Definition 11. Then f H by (15) and (16). Hence f can be considered as a vector in H or a vector in H. Thus H and H can be identified. Definition 14. (Orthogonal) If (u, v) = 0 we write u v and say that u, v are mutually orthogonal. If S, S are subsets of H we say u S if u v, v S (where u H), S S if u v, u S, v S. The set of all u H such that u S is denoted by S. Example 18. u S implies u M where M is the span of S. Let v M, then there are v 1,, v k S and α 1,..., α k C such that v = α 1 v 1 + + α k v k. So thus u S. (u, v) = (u, α 1 v 1 + + α k v k ) = (u, α 1 v 1 ) + + (u, α k v k ) = ᾱ 1 (u, v 1 ) + + ᾱ k (u, v k ) = ᾱ 1 0 + + ᾱ k 0 = 0, Let dim H = N. If x 1,..., x N H have the property (x j, x k ) = δ jk, (54) then they form a basis of H, called an orthonormal basis, for α 1 x 1 +... + α N x N = 0, implies (α 1 x 1 +... + α N x N, x j ) = 0 and α j (x j, x j ) = 0 for all j = 1,..., N, hence α j = 0, showing that x 1,..., x N are linearly independent. 28

2.2 Symmetric operators Definition 15. (Sesquilinear form) Let H, H be two unitary spaces. A complex-valued function t[u, u ] defined for u H and u H is called a sesquilinear form on H H if it is linear in u and semilinear in u. If H = H we speak of a sesquilinear form on H. Let T be a linear operator on H to H, the function t[u, u ] = (Tu, u ) (55) is a sesquilinear form on H H. Conversely, an arbitrary sesquilinear form t[u, u ] on H H can be expressed in this form by a suitable choice of an operator T on H to H. Since t[u, u ] is a semilinear form on H for a fixed u, there exists a unique w H such that t[u, u ] = (w, u ) for all u H. Since w is determined by u, we define a function T by setting w = Tu. T is a linear operator on H to H. In the same way, t[u, u ] can also be expressed in the form t[u, u ] = (u, T u ). (56) Since H, H can be identified with H, H respectively, T can be considered as the adjoint of T on H to H. T T is a linear operator on H to itself. The relation (u, T Tv) = (T Tu, v) = (Tu, Tv) (57) shows that T T is the operator associated with the sesquilinear form (Tu, Tv) on H. Note that the first two members of (57) are the inner product in H while the last is that in H. It follows from (57),(40) that T (Tu, Tv) T = sup sup Tu 2 u v u 2 = T 2. By (41) and (42) we have T T T T = T 2. Hence T T = T 2. (58) Example 19. If T is an operator on H to itself, (Tu, u) = 0 for all u implies T = 0. 29

Proof. We have (T(u+v), u+v) = 0. On the other hand, (T(u+v), u+v) = (Tu, u) + (Tu, v) + (Tv, u) + (Tv, v) = (Tu, v) + (Tv, u) = 0. Since v is any vector in H, then (Tu, iv) + (Tiv, u) = i(tu, v) + i(tv, u) = 0, that is (Tu, v) + (Tv, u) = 0. Combining this equality with (Tu, v) + (Tv, u) = 0 yields (Tv, u) = 0, u, v. Hence T = 0. Remark 3. This property is not true when T is defined on a real space. To show that we can take T represented by the matrix ( ) 0 1 τ jk =, 1 0 then (Tu, u) = 0 for all u, but T 0. Definition 16. A sesquilinear form t[u, v] (or t in short) on a unitary space H is said to be symmetric if t[v, u] = t[u, v], for all u, v H. (59) If t is symmetric, t[u, u] is real-valued, and is denoted by t[u]. t is nonnegative if t[u] 0 for all u, and positive if t[u] > 0 for all u 0. The operator T associated with a symmetric form t[u, v] according to (55),(56) has the property that T = T. (60) Indeed, t[v, u] = (Tv, u) = (v, T u) and t[u, v] = (Tu, v) = (v, Tu). Hence T = T. Definition 17. An operator T on H to itself satisfying (60) is said to be symmetric. (Tu, u) is real for all u H if and only if T is symmetric. Indeed, if (Tu, u) is real, then (Tu, u) = (u, Tu) = (u, Tu) = (T u, u), hence by Example 19 T = T. Conversely if T = T, then (Tu, u) = (T u, u) = (u, T u) = (Tu, u), thus (Tu, u) is real. 30

A symmetric operator T is nonnegative (positive) if the associated form is nonnegative (positive), and we write T 0 to denote that T is nonnegative symmetric. More generally, we write T S or S T if S, T are symmetric operators such that T S 0. Example 20. If T is symmetric, P(T) is symmetric for any polynomial P with real coefficients. Since (αts) = αs T, it follows that T n = T n = T n, thus (α n T n +... + α 0 ) = ᾱ n T n +... + ᾱ 0 = α n T n +... + α 0. Example 21. For any linear operator T on H to H, T T and TT are nonnegative symmetric operators in H and H, respectively. We have (T T) = T (T ) = T T, (T Tu, u) = (Tu, Tu) 0 and (TT ) = (T ) T = TT, (TT u, u ) = (T u, T u ) 0. Example 22. If T is symmetric, then T 2 0; T 2 = 0 T = 0. (T 2 u, u) = (Tu, T u) = (Tu, Tu) 0; T 2 = 0, that is (T 2 u, u) = (Tu, Tu) = 0 according to Example 19, and this is equivalent to T = 0. Example 23. R S and S T imply R T. S T and S T imply S = T. R S and S T is equivalent to S R 0 and T S 0, therefore ((S R)u, u) 0 and ((T S)u, u) 0, hence (Su, u) (Ru, u) 0 and (T u, u) (Su, u) 0. Adding the last two inequalities, we obtain (T u, u) (Ru, u) 0. Hence ((T R)u, u) 0, that is, T R 0. S T and S T is equivalent to S T 0 and S T 0, and this implies that S T = 0, that is (Su, u) (Tu, u) = 0 for all u H. Thus S = T by Example 19. 31

2.3 Unitary, isometric and normal operators Definition 18. Let H and H be unitary spaces. An operator T on H to H is said to be isometric if Tu = u for every u H. (61) This is equivalent to (T Tu, u) = (Tu, Tu) = (u, u), thus This implies that T T = 1. (62) (Tu, Tv) = (u, v) for every u, v H. (63) Definition 19. An isometric operator T is said to be unitary if the range of T is the whole space H. Example 24. T L(H, H ) is unitary if and only if T 1 L(H, H) exists and T 1 = T. (64) If T is unitary, Tu = u implies that the mapping T is one-to-one, that is, T 1 exists. Since T T = 1, T 1 = T. If T 1 exists and T 1 = T, then T T = 1, hence T is unitary. Example 25. T is unitary T is. We have T 1 = T 1 and by Example 24 this is equivalent to (T ) = (T ) 1, which is equivalent to T being unitary. Example 26. If T L(H, H ) and S L(H, H ) are isometric, TS L(H, H ) is isometric. The same is true if isometric is replaced by unitary. We have S S = 1 and T T = 1, therefore (TS) TS = S (T T)S = S S = 1, and if the range of S is the whole space H, and the range of T is the whole space H, this implies that the range of TS is the whole of H. Definition 20. T L(H) is said to be normal if T and T commute, i.e. T T = TT. (65) 32

Symmetric operators and unitary operators on a unitary space into itself are special cases of normal operators. Example 27. Let T 1, T 2, T 3 be operators represented by the matrices 1 1 0 0 0 i 1 i 0 1 1 0, 0 1 0, i 1 0, 0 0 0 i 0 0 0 0 0 respectively, then T 1 is normal, symmetric but not unitary, T 2 is normal, not symmetric but unitary, T 3 is normal, neither symmetric nor unitary. An important property of a normal operator T is that T n = T n, n = 1, 2,... (66) This implies that spr T = T. (67) We shall prove (66). If T is symmetric, by (58) T 2 = T 2. Since T 2 is symmetric we have T 4 = T 2 2 = T 4. Proceeding in this manner we obtain T n = T n for n = 2 m, m = 1, 2,. If T is normal but not necessarily symmetric, again by (58) we have T n 2 = T n T n. Since T n T n = (T T) n because T is normal, and T T is symmetric, by (58) T n 2 = (T T) n = T T n = T 2n for n = 2 m. Now if 2 m n = r 0, then T n T r = T n+r = T n+r T n T r T n T r. Thus T n T n. The opposite inequality is obvious. This proves (66). Example 28. If T is normal then (i) T n = 0 for some integer n implies T = 0. (ii)if T is nonsingular then T 1 is normal. (i) T n = 0 T n = 0 = T n T = 0 T = 0. (ii) We have T T = TT therefore (T T) 1 = (TT ) 1 therefore T 1 T 1 = T 1 T 1, since T 1 = T 1, therefore T 1 T 1 = T 1 T 1. 2.4 Orthogonal projections Definition 21. Let M be a subspace of H and H = M M, the projection operator P = P M on M along M is called the orthogonal projection on M. 33

P is a symmetric, nonnegative and idempotent operator, for (Pu, u) = (u, u + u ) = (u, u ) 0 (68) where u = u + u, u H, u M, u M, (69) by u u. (68) shows also that (Pu, u) is real. Hence P is symmetric, i.e. P = P. Idempotent follows from P being a projection. Example 29. If P is an orthogonal projection, then 1 P is a symmetric, nonnegative and idempotent, and we have 0 P 1, P = 1 if P 0. (1 P) = 1 P = 1 P, ((1 P)u, u) = (u u, u) = (u, u + u ) = (u, u ) 0, and (1 P) 2 = 1 2P + P 2 = 1 2P + P = 1 P. Now ((1 P)u, u) 0, therefore 1 P 0, and (Pu, u) 0, therefore P 0. Hence 0 P 1. Since P 2 = P and P is symmetric, P = P 2 = P 2 by (58), or equivalently P (1 P ) = 0. But P 0 for P 0, therefore P = 1. Example 30. (1 P M )u =dist(u, M), u H. dist(u, M) = inf v M u v. Since u v 2 = (u v ) + u 2 = u v 2 + u 2, the infimum is attained as v = u, and is equal to u = (1 P M )u. Example 31. 1. M N P M P N = 0. 2. The following three conditions are equivalent: (i)m N, (ii)p M P N, (iii)p M P N = P N. Proof. H = M M. (1) We have M N, therefore for each u H, 0 = (P M u, P N u) = (u, P M P Nu) = (u, P M P N u). Hence P M P N = 0. Conversely if P M P N = 0, then P M u = P M P N u = 0 for all u N. Thus M N. (2) Now we will show that (i) (ii), and (i) (iii). If M N, then H = N N 1 M, where N 1 is the orthogonal complement of N in M. Let u = u + u + u, u N, u N 1, u M. 34

(i) (ii): ((P M P N )u, u) = (P M u, u) (P N u, u) = (u + u, u) (u, u) = (u, u) = (u, u ) 0. (ii) (i): Let u N, then (P M u, u) (P N u, u) = (u, u), hence ((I P M )u, u) 0. Since I P M is the orthogonal projection on M, I P M 0 so ((I P M )u, u) = 0 which means that u M. (i) (iii) Since P N u N M, P M P N u = P N u. (iii) (i) Let u N. Then P M P N u = P N u by (iii), and P N u = u. So P M u = u, i.e., u M. 2.5 The eigenvalue problem Example 32. A symmetric operator has only real eigenvalues. We have proved that T = T is equivalent to (Tu, u) being real, and this implies that (λu, u) is real, for an eigenvector u with eigenvalue λ. Thus λ is real. Example 33. Each eigenvalue of a unitary operator has absolute value one. A normal operator with this property is unitary. Since Tu = u u H, Tu = λu implies λu = u, that is λ = 1. T being normal is equivalent to (T Tu, u) = (TT u, u). This is equivalent to (Tu, Tu) = (T u, T u), therefore we have (T u, T u) = (λu, λu) = λ 2 (u, u) = (u, u). On the other hand (T u, T u) = (TT u, u), hence TT = 1. Since T L(H) and is nonsingular, the range of T is the whole space H. Example 34. If T is normal, then we have 1. Tu = 0 if and only if T u = 0. 2. (T λi) is a normal. 3. The eigenvalues of T are the conjugates of the eigenvalues of T. 4. Any two eigenvectors that belong to different eigenvalues are orthogonal. Proof. 1. We will show that (Tu, Tu) = (T u, T u). (Tu, Tu) = (u, T Tu) = (u, TT u) = (T u, T u), so that Tu = 0 is equivalent to T u = 0. 2. We have to show that T λi commutes with its adjoint. (T λi)(t λi) = (T λi)(t λi) = TT λt λt + λ λi = T T λt λt + λ λi = (T λi)(t λi) = (T λi) (T λi), 35

hence T λi is a normal. 3. If Tu = λu, (T λi)u = 0. Since T λi is normal we have (T λi) u = 0, so (T λi)u = 0. Thus T u = λu. Hence any eigenvector of T is also an eigenvector of T, and the corresponding eigenvalues are conjugate to each other. 4. Let Tu 1 = λ 1 u 1, Tu 2 = λ 2 u 2 and λ 1 λ 2, then Since λ 1 λ 2, (u 1, u 2 ) = 0. λ 1 (u 1, u 2 ) = (λ 1 u 1, u 2 ) = (Tu 1, u 2 ) = (u 1, T u 2 ) = (u 1, λ 2 u 2 ) = λ 2 (u 1, u 2 ). 2.6 The minimax principle Let T be a symmetric operator in H. T is diagonalizable (see [2]) and has only real eigenvalues. For a subspace M of H, set (Tu, u) µ[m] = µ[t, M] = min (Tu, u) = min u M 0 u M u 2. (70) u =1 Arrange the eigenvalues of T in the following order µ 1 µ 2... µ N, where µ i can be repeated. The minimax (or maximin) principle is µ n = max µ[m] = max µ[m]. (71) codimm =n 1 codimm n 1 This is equivalent to the following two propositions: µ n µ[m] for any M with codimm n 1; (72) µ n µ[m 0 ] for some M 0 with codimm = n 1. (73) Let us prove these separately. 36

Let {ϕ n } be an orthonormal basis with the property Each u has the expansion u = in this basis. Then Tu = Tϕ n = µ n ϕ n, n = 1,..., N. (74) N ξ n ϕ n, ξ n = (u, ϕ n ), u 2 = n=1 N ξ n Tϕ n = n=1 N µ n ξ n ϕ n, (Tu, u) = n=1 N ξ n 2, n=1 N µ n ξ n 2. Let M be any subspace with codimm n 1. The n-dimensional subspace M spanned by ϕ 1,..., ϕ n contains a nonzero vector u in common with M by (7). This u has the coefficients ξ n+1, ξ n+2,... equal to zero, so that (Tu, u) = = N µ k ξ k 2 1 n µ k ξ k 2 1 n µ n ξ k 2 1 = µ n u 2. (Tu, u) Hence µ[m] = min 0 u M u 2 µ n. This proves (72). Let M 0 be the subspace consisting of all vectors orthogonal to ϕ 1,..., ϕ n 1, so that codimm 0 = n 1. Each u M 0 has the coefficients ξ 1,..., ξ n 1 zero. Hence N (Tu, u) = µ k ξ k 2 which implies (73). k=n N µ n ξ k 2 k=n = µ n u 2, 37 n=1

References [1] T. Kato, Perturbation Theory for Linear Operators, Springer, Berlin, 1976. [2] G. Strang, Linear Algebra and its Applications, Academic Press, New York, 1970. [3] G. Jones, D. Singerman, Complex Functins, University Press, New York, 2002. 38